In complex operational environments, novel fault categories continually emerge in mechanical equipment, whereas available fault samples are severely limited, resulting in degraded diagnostic accuracy. Due to the incremental nature of fault categories and extreme sample scarcity, severe overfitting and catastrophic forgetting are encountered by conventional deep learning models in few-shot class-incremental learning scenarios. To address this challenge, a novel fault diagnosis framework termed Pseudo-Incremental Learning and Inter-class Constraint (PILIC) is proposed in this paper. First, multiple few-shot class-incremental learning pseudo-tasks are constructed from the training dataset, through which the model is guided to capture invariant knowledge shared across base and incremental phases. Second, to resolve feature scale discrepancies caused by sample quantity imbalances between legacy and novel categories, a dynamic calibration module based on the self-attention mechanism is incorporated to minimize the divergence between historical classifiers and emerging class prototypes. Furthermore, an inter-class constraint loss is devised to enlarge the margins between prototypes of distinct fault categories, thereby enhancing the discriminability of feature representations. Finally, the inter-class constraint loss and cross-entropy loss are jointly optimized to collaboratively update the parameters of both the feature extractor and the calibration module, enabling PILIC to reliably classify historical fault categories while novel classes are effectively assimilated. Comprehensive experimental results on motor fault datasets demonstrate that superior diagnostic accuracy is consistently maintained by the proposed method across multiple incremental phases, successfully alleviating overfitting and catastrophic forgetting in few-shot scenarios.